Developing a prediction rule from automated clinical databases to identify high-risk patients in a large population with diabetes

Objective: To develop and validate a prediction rule for identifying diabetic patients at high short-term risk of complications using automated data in a large managed care organization.

Research design and methods: Retrospective cohort analyses were performed in 57,722 diabetic members of Kaiser Permanente, Northern California, aged > or =19 years. Data from 1994 to 1995 were used to model risk for macro- and microvascular complications (n = 3,977), infectious complications (n = 1,580), and metabolic complications (n = 316) during 1996. Candidate predictors (n = 36) included prior inpatient and outpatient diagnoses, laboratory records, pharmacy records, utilization records, and survey data. Using split-sample validation, the risk scores derived from logistic regression models in half of the population were evaluated in the second half. Sensitivity, positive predictive value, and receiver operating characteristics curves were used to compare scores obtained from full models to those derived using simpler approaches.

Results: History of prior complications or related outpatient diagnoses were the strongest predictors in each complications set. For patients without previous events, treatment with insulin alone, serum creatinine > or =1.3 mg/dl, use of two or more antihypertensive medications, HbA(1c) >10%, and albuminuria/microalbuminuria were independent predictors of two or all three complications. Several risk scores derived from multivariate models were more efficient than simply targeting patients with elevated HbA(1c) levels for identifying high-risk patients.

Conclusions: Simple prediction rules based on automated clinical data are useful in planning care management for populations with diabetes.